System protocols, access configurations, and general inquiries.
Yes, Denario is open-source software (GPL-3 License). You can clone the repo and run it locally for free. Note that you will need to provide your own LLM API keys (e.g., OpenAI, Anthropic), which are not free. We also offer a managed cloud version for teams that need high-availability GPUs without setup.
Denario is designed for research labs, PhD students, and autonomous scientists who need to parallelize literature review, hypothesis generation, and simulation coding. It significantly reduces the "blank page" problem in scientific writing.
ChatGPT is a generalist chatbot. Denario is a multi-agent system specifically designed for scientific research. It has access to real-time tools (Python interpreter, plotting libraries), checks against a vast citation database, and employs a "Propose-Critique" loop to self-verify scientific accuracy.
We welcome contributions! Check our GitHub Issues for "good first issue" tags. We are actively looking for help with new tool integrations (e.g., Wolfram Alpha, Mathematica) and additional language support.
Denario works out-of-the-box with GPT-4 and Claude 3.5 Sonnet. Support for local models via Ollama is currently in development and not yet available. You will be able to mix and match models for different agents once released.
To run Denario locally with cloud models (OpenAI/Anthropic), any modern laptop (Mac/Windows/Linux) with Python 3.10+ and Docker is sufficient. To run purely local models (e.g., Llama 3 70B), we recommend at least 32GB of VRAM (e.g., A100 or dual 4090s).
The Peer Review agent is prompted with top-tier conference guidelines (NeurIPS, ICML). It analyzes the drafted paper for clarity, novelty, and methodological soundness, providing a detailed markdown report outlining potential rejection reasons before you submit.
Denario integrates with Semantic Scholar and arXiv APIs. Before finalizing a paragraph, the "Librarian" agent verifies that every cited paper actually exists, matches the claimed findings, and generates a properly formatted BibTeX entry.
Denario is built on a robust Python-based multi-agent architecture. It leverages LangGraph and AG2 for complex agent orchestration and state management, along with CBagent as its specialist deep-research backend. The system integrates with standard scientific Python ecosystem tools (NumPy, SciPy) for rigorous data analysis.
If you choose to use proprietary models like GPT-4o or Claude 3.5 Sonnet, you will need to provide your own API keys in the `.env` file. These keys remain on your machine. If you use local models via Ollama/vLLM, no keys are required.
Yes. Executing code in Docker containers is supported and highly recommended for safety, though it is not compulsory. The system has no network access unless explicitly whitelisted, ensuring your local environment remains safe from generated code side-effects.
No. When you run Denario locally, your data never leaves your machine. For our managed cloud service, we have a strict zero-retention policy and enterprise agreements (BAA) available for healthcare applications.
Like all LLM-based systems, raw generation can hallucinate. However, Denario mitigates this via a "Librarian" agent that cross-references every claim against real-time APIs (Semantic Scholar, arXiv). If a citation cannot be verified, the "Critic" agent rejects the paragraph. We also strictly separate the "Creative" agents (who hypothesize) from the "Coder" agents (who must run valid executable python).
Transparency is key: currently, we estimate about 1 in 10 generated insights are truly novel and actionable. The other 90% tend to be confirmatory or incremental. Denario is a tool to accelerate the discovery of that 10%, not a magic button that guarantees a Nobel Prize on every click.
Denario operates in a "Code-Execution Loop". It doesn't just guess code; it writes, runs, captures the stdout/stderr, debugging its own errors until the script executes successfully. However, human review is mandatory. You should treat Denario as a junior research assistant: capable of great work, but requiring supervision and final validation.
